On Mon, Jun 22, 2015 at 5:04 PM, NeilBrown <neilb@xxxxxxxx> wrote: > > On Mon, 22 Jun 2015 07:41:11 -0400 > Trond Myklebust <trond.myklebust@xxxxxxxxxxxxxxx> wrote: > > > Hi Neil, > > > > On Mon, Jun 22, 2015 at 3:53 AM, NeilBrown <neilb@xxxxxxx> wrote: > > > > > > Hi, > > > this is my proposed solution to the problem I outlined in > > > NFSv4 state management issue - Linux disagrees with Netapp. > > > I haven't tested it yet (no direct access to the Netapp), but > > > I'll try to get some testing done. RFC for now. > > > > > > NeilBrown > > > > > > > > > When opening a file, nfs _nfs4_do_open() will return any > > > incompatible delegation, meaning if the delegation held for > > > that file does not give all the permissions required, it is > > > returned. > > > This is because various places assume that the current delegation > > > provides all necessary access. > > > > > > However when a delegation is received, it is not validated in the > > > same way so it is possible to, for example, hold a read-only > > > delegation while the file is open write-only. > > > When that delegation is recalled, the NFS client will try to > > > reclaim the write-only open, and that will fail. > > > > > > > I'd argue that the bug here is the attempt to reclaim the write-only > > open; your previous email appeared to show that the client already > > held a corresponding open stateid. > > I did consider that approach, but I managed to talk myself out of it... > Let's see if I can talk you out of it too. > > There are potentially two state ids available for each open_owner+inode > - an open_stateid and a delegation stateid. > > Linux does track which of read/write the delegation stateid permits, > but does *not* track which the open_stateid permits. > So when returning a delegation it does not know which of "read" and > "write" need to be reclaimed (because open_stateid doesn't provide > them) but it does know which cannot be reclaimed (because delegation > stateid didn't provide them) - so it could just reclaim whatever it > needs that the delegation *could* have provided. > So this particular bug could be fixed that way. > > However, consider the scenario I described up to just before the 'link' > system call. > The client holds a write-only open_stateid and a read-only delegation > stateid. > If the client (same lockowner) opens the file read-only again the open > will succeed without talking to the server on the strength of the > delegation. > update_open_stateid will then copy the delegation stateid into the state > and all IO will use that stateid. If a write is attempted with the > still-open write-only fd, it will use the read-only delegation stateid > and presumably get an error. This is incorrect. As far as I know, a 4.1 client will do the following: The NFSv4 open() code will catch the delegation as being insufficient using can_open_delegated(), and will ensure that the client calls OPEN in this case. The resulting open stateid is then saved in the state->open_stateid. If an I/O attempt is then made for an I/O type for which the delegation cannot be used, then nfs4_select_rw_stateid() will return either the lock stateid or the open stateid; whichever is appropriate. > Unless I've missed something there is no code in Linux/NFS to > selectively use one stateid for reads and another for writes - both > coming from the same lockowner to the same inode. See above. > Presumably this is the reason that we have > nf4_return_incompatible_delegation(): because Linux/NFS assumes that if > it holds a delegation, that delegation covers all active open modes. > For exactly the same reason, we need to reject a delegation if it > doesn't cover all the open modes that are already active. > > Certainly we *could* track exactly which accesses the open_stateid > allows, and could have (potentially) separate "read" and "write" > stateids, but that paths wasn't the easiest so I didn't follow it. > I'm rather thinking that the simplest fix is simply to have nfs4_open_delegation_recall() skip those file modes for which the current delegation stateid is not appropriate. From a client perspective, that should always make sense. Cheers Trond -- To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
